automatic classiﬁcation of calciﬁcation in the coronary ... faculteit/decaan... · automatic...

Automatic Classification of Calcification in the

Coronary Vessel Tree

R. Shahzad1,2,3, L. J. van Vliet2, W. J. Niessen2,3, and T. van Walsum3

1 Division of Image Processing, Department of Radiology,Leiden University Medical Center, Leiden, The Netherlands2 Quantitative Imaging Group, Faculty of Applied Science,

Delft University of Technology, Delft, The Netherlands3 Biomedical Imaging Group Rotterdam, Department of Radiology,

Erasmus Medical Center, Rotterdam, The Netherlands

Abstract. This paper presents a pipeline for complete automaticdetection and labelling of the coronary artery calcium lesions. Thepipeline not only distinguishes between coronary artery calcificationsand non-calcifications but also labels the detected calcium lesion toone of the main coronary vessel tree. The pipeline presented uses acombination of atlas-based and machine learning approaches. The setof features supplied to the classifier are either directly obtained from thesubject scans or are derived after using information from atlas scans. Theclassifier was trained using 76 CT datasets. The pipeline was tuned usingthe training set provided by the challenge evaluation framework. On the32 test datasets we achieved a sensitivity of 67% and a PPV of 85%,comparing the volume of the detected calcifications it was 89% and 94%respectively. The ICC was 0.99 for LAD+LM, 0.98 for LCX and 0.92 forRCA, on the whole heart the ICC for Agatston and volume score were0.99 and 0.98 respectively. An accuracy of 88% and a linear weightedkappa (κ) of 0.90 was achieved for the risk categorization accuracy.

Keywords: Coronary artery, calcium scoring, atlas, CT, CTA

1 Introduction

Coronary artery disease (CAD) is one of the leading causes of mortalityworldwide [1]. Many clinical studies have shown that the amount of calciumin coronary artery plaques correlates with the risk of future cardiovascularevents [2, 3]. It has also been suggested that vessel-specific calcium scoring ismore informative compared to whole heart calcium scoring [4].

Manual calcium scoring is a time consuming task because it consists ofdrawing contours to obtain the region of interest or clicking inside all calciumobjects. Lately, a lot of interest has been shown in automating the task ofcalcium detection/scoring on CT scans. Isgum et al [5] and Kurkure et al [6]have demonstrated the feasibility of automating the task for whole heart calcium

2 Automatic Calcium Lesion Detection

scoring. Shahzad et al [7] have demonstrated the feasibility of automatic vessel-specific calcium scoring. A few other studies also perform detection/scoring ofcalcium lesions on Computed Tomography angiography (CTA) scans [8–11].

In this paper we present a pipeline that enables us to automatically detectvessel-specific calcium lesions, and evaluate it using the orCaScore evaluationframework. The pipeline is an adaptation of our earlier work [7]. The mainchange is an improved initialization of the atlas registration.

2 Method

The method used is similar to [7], with some modifications. The changes tothe method is the initialization of the atlas registrations to be able to work onmulti-vendor datasets. The method consists of the following steps. 1) Performa multi-atlas based registration, 2) detect candidate calcium lesions from theCT scans, 3) a set of image features are computed for the detected candidatelesions, 4) a trained classifier labels the candidate lesions either as a calcificationor noise, and 5) the lesions with the calcification label gets assigned to one ofthe coronary vessels viz. left anterior descending + left main (LAD+LM), leftcircumflex (LCX) and right coronary artery (RCA).

2.1 Atlas registration

A multi-atlas based registration is used in our method. The atlas registrationhas two purposes: 1) to segment the heart from the CTA scans [12], and 2) tocompute a set location based features for the classifier.

We use a set of eight atlas CTA scans. All the atlas scans have beenacquired on the same machine (Somatom Sensation, Siemens Medical Solutions,Forchheim, Germany). Each of the atlases has a segmentation of the heart[13], coronary artery location estimate for each individual artery [14], and astandardized coordinate space [7].

ElastiX, a publicly available registration package [15] is used to registerthe atlas CTA scans to the subjects CTA scan. Registration is performed intwo stages using a multi-resolution strategy. An initial rigid registration isused to approximately align the two scans. This rigid registration is followedby a non-rigid registration. The rigid registration is performed on maskeddistance transform scans using normalized cross correlation as a similaritymeasure. The non-rigid registration is performed on the original CTA scans,a B-spline transformation model was employed while using the result of therigid transformation to initialize the registration, mutual information is used asthe similarity measure. The resulting transformations from the registration stepsare used to map the atlas heart segmentation, coronary artery location estimatesand the standardized coordinate space onto the subject scan. See Figure 1 foran overview of the registration stage. Finally, the information from all the eightatlases are combined (majority voting for the heart segmentation, averagingfor coronary artery location estimates and standardized coordinate space) and

Automatic Calcium Lesion Detection 3

mapped onto the subject’s CT scan using an additional intra-subject (CTA-CT)registration step.

Atlas Scan

Subject Scan

Multi-Resolution

Multi-Stage

Registration

Transformations

Atlas Information

Standardized

coordinate spaceCoronary artery

location estimate

Whole-Heart

segmentation

Labels

Fig. 1. An overview of the registration stage, this stage is performed for each of theeight atlas scans.

2.2 Candidate detection

Candidate calcium lesions are obtained from the CT scans by thresholding thesegmented heart at 130 HU (See Figure 2). All objects with a size smaller than1.5 mm3 are discarded, as these are assumed to be noise. The detected candidatelesions consist of true coronary calcifications, aortic calcifications, calcificationsin the valves, and false objects due to noise and imaging artefacts.

2.3 Feature Computation

For each of the detected candidate lesions a set of 21 features are calculated (seeTable 1 for the detailed list). The features are divided into subcategories:

Object based features - Three object based features are computed. 1) volume,2) maximum intensity and, 3) average intensity of the candidate lesion.

Multi-scale image derivatives - The intensity of the object at the maximum


Fig. 2. A random axial slice from one of the CT training dataset, the blue overlay showsthe segmented heart, the red overlay represents the coronary artery density estimates(obtained after the atlas-based segmentations) and the green objects represents thedetected candidate lesions.

intensity point after Gaussian image derivatives were computed at five differentscales (Gaussian standard deviation in-slice between 0.3 mm to 4.8 mm withone sample per octave; between slices from 1.5 mm to 24 mm with one sampleper octave) up to the second order. A total of 15 multi-scale image derivativefeatures are computed.

Location estimates - The positions of the candidates in the actual image spaceand standardized coordinate space in the z-axis are computed. The coronaryartery location feature is also computed. Resulting in three features.

2.4 Classifer

A 9-nearest neighbour (NN) classifier is used to help us distinguish thethe candidate objects. The classifier was trained on 76 CT datasets thatwere obtained from a Siemens scanner (Definition, Siemens Medical Solutions,Forchheim, Germany). More details about the scans and the classifier trainingcan be obtained from [7].

3 Results

Our method was applied on both the training and the test datasets which wereprovided by the orCaScore challenge organizers. Table 2 shows the results of ourpipeline on the training datasets for each of the vendors.


Table 1. Set of features used.

Feature Description Number of features

Volume of object in mm3 1

Intensity of object: maximum and average 2

Position: z coordinate in the image and mean space 2

Coronary artery location estimate 1

Intensity of voxel after 4

Gaussian filtering at scale 1, 2, 4, 8

Intensity of voxel after 1st 3

order Gaussian derative in z direction at scale 1, 8, 16

Intensity of voxel after 2nd 6

order Gaussian deratives in xx, yy direction at scale 1, 2, 4

Intensity of voxel after 2nd 2

order Gaussian deratives in zz direction at scale 1, 2

Table 2. Performance of the pipeline on the training datasets.

Vendor Id 1 2 3 4 Average

Sensitivity (%) 83 80 68 68 75Specificity (%) 99 99 98 99 99PVV (%) 78 77 50 78 71Volume TP Detected (%) 74 92 84 92 86Avg no. FN 0.87 1.12 0.87 2.65 1.38Avg no. FP 1.87 3.25 2.50 2.65 2.56Avg FP volume mm3 11.7 129.4 41.8 53.1 46.4

On the 32 testing datasets we achieved an overall sensitivity of 67% anda positive predictive value (PPV) of 85%, with respect to the volume of theidentified calcifications we were able to detect 89% and 94% respectively. Theinterclass correlation coefficient (ICC) for Agatston score and volume score was0.99 and 0.98 respectively. For vessel specific volume scores we obtained an ICCof 0.99 for LAD+LM, 0.98 for LCX and 0.92 for RCA. Comparing the riskcategory accuracy of our automatic method to the manual Agatston we obtaina correlation of 0.88 and a linearly weighted kappa (κ) value of 0.90.

4 Discussion and Conclusions

The results on the testing datasets show a sensitivity of 67% and PPV of 85%,whereas 90% of the the volume of the calcium lesions have been detected.This shows that the lesions that were missed were small calcium lesions thatcould be easily mistaken as noise. When we compare the risk categorization


(a) (b)

(c)

Fig. 3. The figures show a few axial slices from one of the test datasets after applyingour pipeline. Pink overlay represents the structures with intensities above 130 HU, redobjects are labelled as lesions belonging to LAD+LM, yellow to LCX and green toRCA

accuracy, there is very close agreement with the manually obtained ground truth.It can also be appreciated that our atlas-based method can be easily appliedon data from different vendors or different scanning protocols without loosingperformance. It can be noted from the training datasets that the performanceon scans acquired on different scanners is similar. Also, from the training data itwas observed that our method is sometimes prone to registration related errors,especially when the field of view (FOV) of the subjects scan was very differentfrom the atlases used. In the present pipeline we do not use an extensive atlascombination/selection strategy, but we simply average the labels together. Thissometimes causes FP or FN detections. The method described was also evaluatedon another dataset consisting of 56 datasets from a Siemens scanner [7], onthis dataset we achieved a sensitivity of 87%, with respect to risk categoryclassification 89% of the subjects were assigned to the correct risk category.


We obtained vessel-specific correlations of 0.99, 0.93 and 0.93 for the LAD+LM,LCX and RCA volume scores respectively.

In the current method both CTA and CT scans were used. In case onlyCT scans are available for analysis, our method can easily be adapted to workonly on the CT scans with some modifications to the registration parameters.The registrations and transformations are the most computationally expensiveprocess in our pipeline, this takes approximately 10 minutes for each of theatlases, feature set computation and classification takes a couple of seconds ona Intel Xeon CPU [email protected] with 16GB RAM. This process can beparallelized further by running the proposed pipeline on a computational cluster.

In conclusion, we presented an adapted version of our earlier calcium scoringmethod, which was based on a combination of atlas-based and machine learningapproaches. The method was evaluated on the orCaScore framework, andobtained a PPV of 85%, an ICC of 0.99 and 0.98 for Agatston and volume scorerespectively, vessel-specific correlation of 0.99, 0.98 and 0.92 for LAD+LM, LCXand RCA respectively. With respect to risk category classification we obtaineda correlation of 0.88 and a κ value of 0.90.

Acknowledgements

We would like to thank Dr. Hortense Kirisli and Dr. Nora Baka for theirsuggestions and fruitful discussions.

References

1. WHO: Cardiovascular diseases, fact sheet 317. World Health Organization. (2013)2. Budoff, M.J., Gul, K.M.: Expert review on coronary calcium. Vascular health and

risk management 4(2) (2008) 3153. Elias-Smale, S.E., Proenca, R.V., Koller, M.T., et al.: Coronary calcium score

improves classification of coronary heart disease risk in the elderly: The RotterdamStudy. Journal of the American College of Cardiology 56(17) (2010) 1407–1414

4. Rosen, B.D., Fernandes, V., McClelland, R.L., et al.: Relationship between baselinecoronary calcium score and demonstration of coronary artery stenoses duringfollow-up: MESA (multi-ethnic study of atherosclerosis). JACC: CardiovascularImaging 2(10) (2009) 1175–1183

5. Isgum, I., Prokop, M., Niemeijer, M., Viergever, M.A., van Ginneken, B.:Automatic coronary calcium scoring in low-dose chest computed tomography.Medical Imaging, IEEE Transactions on 31(12) (2012) 2322–2334

6. Kurkure, U., Chittajallu, D.R., Brunner, G., Le, Y.H., Kakadiaris, I.A.: Asupervised classification-based method for coronary calcium detection in non-contrast CT. The international journal of cardiovascular imaging 26(7) (2010)817–828

7. Shahzad, R., van Walsum, T., Schaap, M., et al.: Vessel specific coronary arterycalcium scoring: An automatic system. Academic radiology 20(1) (2013) 1–9

8. Ebersberger, U., Eilot, D., Goldenberg, R., et al.: Fully automated derivation ofcoronary artery calcium scores and cardiovascular risk assessment from contrastmedium-enhanced coronary CT angiography studies. European radiology 23(3)(2013) 650–657


9. Rubinshtein, R., Halon, D.A., Gaspar, T., Lewis, B.S., Peled, N.: Automaticassessment of coronary artery calcium score from contrast-enhanced 256-rowcoronary computed tomography angiography. The American journal of cardiology113(1) (2014) 7–11

10. Eilot, D., Goldenberg, R.: Fully automatic model-based calcium segmentation andscoring in coronary CT angiography. International journal of computer assistedradiology and surgery (2013) 1–14

11. de Graaf, M.A., Broersen, A., Ahmed, W., et al.: Feasibility of an automatedquantitative computed tomography angiography–derived risk score for riskstratification of patients with suspected coronary artery disease. The Americanjournal of cardiology 113(12) (2014) 1947–1955

12. Kirisli, H., Schaap, M., Klein, S., et al.: Evaluation of a multi-atlas based methodfor segmentation of cardiac CTA data: a large-scale, multicenter, and multivendorstudy. Medical physics 37(12) (2010) 6279–6291

13. Shahzad, R., Bos, D., Metz, C., et al.: Automatic quantification of epicardialfat volume on non-enhanced cardiac CT scans using a multi-atlas segmentationapproach. Medical physics 40(9) (2013) 091910

14. Shahzad, R., Schaap, M., van Walsum, T., et al.: A patient-specific coronarydensity estimate. In: IEEE International Symposium on Biomedical Imaging.(2010) 9–12

15. Klein, S., Staring, M., Murphy, K., Viergever, M., Pluim, J.: elastix: a toolbox forintensity-based medical image registration. Medical Imaging, IEEE Transactionson 29(1) (2010) 196–205

automatic classiﬁcation of calciﬁcation in the coronary ... faculteit/decaan... · automatic...

Documents